Add more options to fail on insufficient coverage

[michaelvlach]

Enforcing code coverage is good and the current option --fail-under-lines is a way to do it but unfortunately rather crude one. Percentage is relative and may change due to amount of code changing vis-a-vis covered code. The llvm-cov report does provide absolute values of covered and uncovered for each of the 4 categories:

• function
• line
• region
• branch

The better way would therefore be to fail coverage on uncovered absolute values rather than percentages which is volatile. That way if new code is added that is fully covered the coverage will pass. However when new code is not fully covered it will also require changing of the threshold(s) which is good thing - it might be justified of course but nevertheless should be visible.

I propose the following options:

--fail-uncovered-functions
--fail-uncovered-lines
--fail-uncovered-branches
--fail-uncovered-regions

[taiki-e]

I think referencing absolute values instead of percentages is an interesting idea. I'm fine with adding this as an experimental option.

However when new code is not fully covered it will also require changing of the threshold(s) which is good thing

I think this is not necessarily a good thing when there are multiple PRs, because it easily causes merge conflicts.

[michaelvlach]

I think referencing absolute values instead of percentages is an interesting idea. I'm fine with adding this as an experimental option.

Thanks!

However when new code is not fully covered it will also require changing of the threshold(s) which is good thing

I think this is not necessarily a good thing when there are multiple PRs, because it easily causes merge conflicts.

This would happen only if you had multiple PRs that are all worsening coverage. Something nobody would want I presume. And if it happens regularly then enforcing coverage is probably not what is wanted.

[taiki-e]

This would happen only if you had multiple PRs that are all worsening coverage.

Note that this can also be caused by PRs that improve coverage (if the current coverage is not 100%).

• PR A adds a test for function A1 and reduces uncovered lines from 5 to 3.
• PR B adds a test for function B1 and reduces uncovered lines from 5 to 4.
• Merging both PRs reduces uncovered lines from 5 to 2, but there is a merge conflict.

EDIT: see #170 (comment)

[taiki-e]

Ah, sorry, I somehow misunderstood that the values passed to those options are concrete values. It is thresholds, so my comment above is wrong.